Virtual workplace training task allocation method and system based on multi-agent collaboration

Abstract

This invention provides a virtual workplace training task allocation method and system based on multi-agent collaboration, relating to the field of intelligent training technology. The method involves decomposing training tasks into subtasks by constructing a knowledge dependency graph; calculating scores for agent knowledge complementarity, coordination proficiency, and task-carrying capacity, and integrating them into a collaborative score; establishing a knowledge topology network based on difficulty coefficients and collaborative scores to select the optimal agent combination; and recording training data to optimize the allocation mechanism. This invention improves training efficiency, achieves optimal allocation of agent resources, and enhances the targetedness and adaptability of training.

Description

Technical Field

[0001] The present invention relates to intelligent training technology, and in particular to a virtual workplace training task allocation method and system based on multi-agent collaboration. Background Art

[0002] Virtual workplace training systems, which use agents to collaborate and complete training tasks, have become an important means of improving corporate training efficiency. Existing virtual workplace training methods primarily rely on a single agent to impart knowledge, which struggles to meet the multi-dimensional requirements of complex training tasks and lacks systematic evaluation of the agent's knowledge coverage and collaborative capabilities.

[0003] As training tasks become increasingly complex, single agents face bottlenecks in both knowledge breadth and processing power. While multi-agent collaborative training methods are gaining popularity, existing approaches suffer from issues like irrational task decomposition, blind agent selection, and low collaborative efficiency, making it difficult to fully leverage the strengths of a swarm of agents.

[0004] Therefore, there is an urgent need for a multi-agent collaborative training task allocation method based on a knowledge topological network. By analyzing the knowledge dependencies of training tasks, tasks are decomposed, and the collaborative effects between agents are evaluated to establish a knowledge transfer network, dynamic optimization allocation of training tasks is achieved, and the effectiveness of virtual workplace training is improved. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for allocating virtual workplace training tasks based on multi-agent collaboration, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for allocating virtual workplace training tasks based on multi-agent collaboration, comprising:

[0007] Collect training task requirements and the operating status of the virtual workplace training agent, build a knowledge dependency graph based on the training task requirements, identify the hierarchical relationship and degree of correlation between training knowledge nodes, and divide the training task into multiple training subtasks; establish an execution priority sequence for the training subtasks, evaluate the difficulty coefficient of each training subtask based on the knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating;

[0008] Based on the running status of the virtual workplace training agents, the knowledge complementarity score, cooperation proficiency score and task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score;

[0009] Based on the difficulty coefficient and agent collaboration scores in the training task decomposition scheme, a knowledge topology network is established. The transfer path scores of the agents in the knowledge topology network are extracted. The agent with the highest transfer path score and whose computing resources are not saturated is selected as the leading training agent. The collaborative training agents are determined based on the knowledge distribution of the transfer path, and the agent combination configuration scheme for the training subtasks is generated.

[0010] Execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

[0011] In an optional embodiment,

[0012] Based on the training task requirements, a knowledge dependency graph is constructed to identify the hierarchical relationship and degree of association between training knowledge nodes. The training task is then divided into multiple training subtasks, including:

[0013] Extracting knowledge points from training task requirements, converting the knowledge points into vector form, calculating the distance between the vectors of any two knowledge points to obtain a dependency strength value, and constructing a knowledge dependency matrix based on the dependency strength value;

[0014] Performing singular value decomposition on the knowledge dependency matrix to obtain eigenvectors of the knowledge points, using the eigenvectors to construct a distribution structure of the knowledge points in the knowledge space, and generating a knowledge dependency graph;

[0015] In the knowledge dependency graph, the out-degree value and the in-degree value of each knowledge point are calculated, and the hierarchical position coefficient of the current knowledge point is determined according to the ratio of the out-degree value to the in-degree value;

[0016] The knowledge points in the knowledge dependency graph are layered according to the hierarchical position coefficients, and the knowledge points with the same hierarchical position coefficients and dependency strength values ​​greater than a preset strength threshold are divided into a group to obtain a plurality of training subtasks.

[0017] In an optional embodiment,

[0018] Establish an execution priority sequence for training subtasks, evaluate the difficulty coefficient of each training subtask based on knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating.

[0019] Obtain the knowledge dependency relationship of the training subtask, calculate the number of direct input edges and direct output edges of the knowledge node to obtain the initial dependency value, calculate the number of indirect input edges and indirect output edges of the knowledge node to obtain the transitive dependency value, and calculate the dependency coefficient of the training subtask based on the initial dependency value and the transitive dependency value;

[0020] Marking reachable nodes on the knowledge dependency graph, calculating a node transfer factor according to the hierarchical distribution of the reachable nodes, and taking the product of the node transfer factor and the total number of reachable nodes as the importance value of the training subtask;

[0021] A state transition matrix is ​​constructed using the dependency coefficients and importance values, a state distribution vector is calculated by eigenvalue decomposition, the state distribution vector is used as the execution priority sequence of the training subtasks, the degree centrality and betweenness centrality of the knowledge nodes are calculated, the weighted sum of the degree centrality and betweenness centrality is used as the node coverage, and the node weight is calculated in combination with the reference chain length of the knowledge node;

[0022] The weights of all knowledge nodes in the training subtasks are accumulated to obtain a difficulty coefficient, and the difficulty coefficients of all training subtasks are graded according to the execution priority sequence to generate a training task decomposition plan including an execution order and a difficulty rating.

[0023] In an optional embodiment,

[0024] Marking reachable nodes on the knowledge dependency graph and calculating the node transfer factor according to the hierarchical distribution of the reachable nodes include:

[0025] Selecting a starting node in a knowledge dependency graph, generating a node access mark sequence, traversing the knowledge dependency graph using a depth-first search, recording the node access depth and access path, marking the target node that can be reached by the starting node as a reachable node, and forming an access record of the reachable node;

[0026] Dividing the node hierarchy by using the access depth of the node in the access record, calculating the dependency weights of adjacent nodes, assigning hierarchical identifiers to the reachable nodes based on the dependency weights and the access depth, and obtaining a hierarchical distribution structure of the nodes;

[0027] Counting the number of preceding and succeeding nodes of the nodes based on the hierarchical distribution structure, calculating the rate of change of the number of nodes in adjacent levels, and calculating the inter-layer connection strength of the nodes based on the rate of change and the number of preceding and succeeding nodes;

[0028] A node diffusion factor is generated based on the inter-layer connection strength, and a node transfer probability table is constructed in combination with the hierarchical distribution structure. The hierarchical influence strength of the node is calculated through the transfer probability table. The node diffusion factor and the hierarchical influence strength are weightedly combined to generate a node knowledge transfer factor.

[0029] In an optional embodiment,

[0030] Based on the running status of the virtual workplace training agent, the knowledge complementarity score, cooperation proficiency score and task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score, including:

[0031] Obtaining the operational status data of the virtual workplace training agent, including the agent's knowledge status data, interaction status data, and resource status data;

[0032] Extracting the knowledge vectors and skill vectors of the agents based on the knowledge state data, calculating the difference distribution of the knowledge structures between the agents, analyzing the coverage of the knowledge transfer paths between the agents, and calculating the knowledge complementarity scores between the agents based on the difference distribution and coverage;

[0033] Extracting collaborative records of agents based on the interaction state data, including interaction delay, interaction duration, and task achievement, introducing a time decay function into the collaborative records, and calculating a cooperation proficiency score between agents based on the collaborative records processed by the time decay function;

[0034] Analyze the computing resource occupancy, storage resource occupancy, and communication resource occupancy of the agent based on the resource status data, evaluate the resource scheduling efficiency and task switching loss of the agent, and calculate the task carrying capacity score of the agent based on the resource scheduling efficiency and task switching loss;

[0035] The task scenario features are extracted, and the weight coefficients of each score are calculated based on the task scenario features. The knowledge complementarity score, cooperation proficiency score, and task carrying capacity score are nonlinearly weighted and fused with the weight coefficients to generate an intelligent agent collaborative score.

[0036] In an optional embodiment,

[0037] Based on the difficulty coefficient in the training task decomposition plan and the agent collaborative score, a knowledge topology network is established. The transfer path scores of the agents in the knowledge topology network are extracted. The agents with the highest transfer path scores and whose computing resources are not saturated are selected as the leading training agents.

[0038] Extracting the difficulty coefficient of the training subtask from the training task decomposition plan, converting the difficulty coefficient into a node-based weight, and converting the agent collaboration score into a node-dynamic weight;

[0039] An initial knowledge topology network is constructed using the node basic weights and node dynamic weights, the knowledge hierarchy relationship of adjacent nodes in the initial knowledge topology network is analyzed, the knowledge overlap between nodes is calculated, the direction of knowledge flow is determined based on the knowledge hierarchy relationship and the knowledge overlap, and a directed transfer path is constructed;

[0040] Calculating the node complexity on the directed transfer path based on the directed transfer path, substituting the node complexity and the transfer distance into an exponential decay function to obtain a path knowledge decay value, and multiplying the knowledge overlap between nodes by the path knowledge decay value to obtain a transfer path score;

[0041] The task queue length, memory occupancy rate and response delay time of the intelligent agent are collected, and the processing load index is calculated. The processing load index is compared with the preset multi-dimensional saturation threshold to obtain the resource status judgment result of the intelligent agent. The transfer path score and the resource status judgment result are combined and evaluated, and the intelligent agent with the highest transfer path score and which has not reached the resource saturation state is selected as the leading training intelligent agent.

[0042] In an optional embodiment,

[0043] The collaborative training agent is determined based on the knowledge distribution of the transfer path, and the agent combination configuration scheme for generating training subtasks includes:

[0044] Extract the transfer path of the leading training agent in the knowledge topology network, calculate the knowledge weight of each node on the path, build a knowledge transfer chain based on the knowledge association strength between nodes, analyze the knowledge coverage based on the knowledge transfer chain, and generate knowledge distribution characteristics;

[0045] Calculate the matching degree between the knowledge distribution characteristics and the training subtask requirements, identify uncovered knowledge content, search for agents with the uncovered knowledge content in the knowledge topology network, calculate the degree of knowledge complementarity between the candidate agent and the leading training agent based on the knowledge association strength, and select the agent with the highest degree of knowledge complementarity as the collaborative training agent;

[0046] Based on the strength of the knowledge association, the order of knowledge transfer between the leading training agent and the collaborative training agent is determined, computing resources are allocated to the agents according to the knowledge distribution characteristics and the degree of knowledge complementarity, and the knowledge transfer order and computing resource allocation results are combined to generate an agent combination configuration plan for the training subtask.

[0047] A second aspect of an embodiment of the present invention provides a virtual workplace training task allocation system based on multi-agent collaboration, comprising:

[0048] The first unit is used to collect training task requirements and the operating status of the virtual workplace training agent. Based on the training task requirements, it constructs a knowledge dependency graph, identifies the hierarchical relationships and correlations between training knowledge nodes, and divides the training task into multiple training subtasks. It also establishes an execution priority sequence for the training subtasks, assesses the difficulty coefficient of each training subtask based on the knowledge node coverage, and generates a training task decomposition plan that includes the execution order and difficulty rating.

[0049] The second unit is used to calculate the knowledge complementarity score, cooperation proficiency score, and task carrying capacity score between agents based on the operating status of the virtual workplace training agents, and to weight and fuse these scores to generate an agent collaboration score;

[0050] The third unit is used to establish a knowledge topology network based on the difficulty coefficient and agent collaboration score in the training task decomposition plan, extract the transfer path scores of the agents in the knowledge topology network, select the agent with the highest transfer path score and whose computing resources are not saturated as the leading training agent, and determine the collaborative training agents based on the knowledge distribution of the transfer path to generate the agent combination configuration plan for the training subtasks;

[0051] The fourth unit is used to execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

[0052] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0057] In this embodiment, a virtual workplace training task allocation method based on multi-agent collaboration is used to achieve accurate decomposition and intelligent allocation of training tasks, thereby improving the efficiency and quality of virtual workplace training. By constructing a knowledge dependency graph and a knowledge topology network, it is possible to scientifically evaluate the difficulty coefficient of training tasks and rationally allocate agent resources, ensuring the systematicness and coherence of the training content while fully utilizing the knowledge complementarity and collaborative advantages of different agents. The present invention also establishes a dynamic feedback optimization mechanism, which continuously updates the agent collaborative score by recording the training process data, so that the task allocation process can be adaptively adjusted, thereby achieving continuous improvement in training effectiveness and efficient use of system resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a method for allocating virtual workplace training tasks based on multi-agent collaboration according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0060] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG is a flow chart of a method for allocating virtual workplace training tasks based on multi-agent collaboration according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Collect training task requirements and the operating status of the virtual workplace training agent, build a knowledge dependency graph based on the training task requirements, identify the hierarchical relationship and degree of correlation between training knowledge nodes, and divide the training task into multiple training subtasks; establish an execution priority sequence for the training subtasks, evaluate the difficulty coefficient of each training subtask based on the knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating;

[0063] Based on the running status of the virtual workplace training agents, the knowledge complementarity score, cooperation proficiency score and task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score;

[0064] Based on the difficulty coefficient and agent collaboration scores in the training task decomposition scheme, a knowledge topology network is established. The transfer path scores of the agents in the knowledge topology network are extracted. The agent with the highest transfer path score and whose computing resources are not saturated is selected as the leading training agent. The collaborative training agents are determined based on the knowledge distribution of the transfer path, and the agent combination configuration scheme for the training subtasks is generated.

[0065] Execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

[0066] In an optional embodiment, a knowledge dependency graph is constructed based on the training task requirements, the hierarchical relationship and correlation between training knowledge nodes are identified, and the training task is divided into multiple training subtasks including:

[0067] Extracting knowledge points from training task requirements, converting the knowledge points into vector form, calculating the distance between the vectors of any two knowledge points to obtain a dependency strength value, and constructing a knowledge dependency matrix based on the dependency strength value;

[0068] Performing singular value decomposition on the knowledge dependency matrix to obtain eigenvectors of the knowledge points, using the eigenvectors to construct a distribution structure of the knowledge points in the knowledge space, and generating a knowledge dependency graph;

[0069] In the knowledge dependency graph, the out-degree value and the in-degree value of each knowledge point are calculated, and the hierarchical position coefficient of the current knowledge point is determined according to the ratio of the out-degree value to the in-degree value;

[0070] The knowledge points in the knowledge dependency graph are layered according to the hierarchical position coefficients, and the knowledge points with the same hierarchical position coefficients and dependency strength values ​​greater than a preset strength threshold are divided into a group to obtain a plurality of training subtasks.

[0071] This embodiment provides a method for constructing a knowledge dependency graph based on training task requirements. The method can identify the hierarchical relationship and correlation degree between training knowledge nodes, divide the training task into multiple training subtasks, thereby improving training efficiency and learning effects.

[0072] In this implementation, extracting knowledge points from training task requirements is a fundamental step in building a knowledge dependency graph. For example, for a task requirement called "Basic Machine Learning Training," the system uses text analysis techniques to extract relevant knowledge points, such as "linear regression," "decision tree," "neural network," "data preprocessing," and "model evaluation." After extraction, the system uses natural language processing techniques to convert these knowledge points into vector form, with each knowledge point represented as a high-dimensional vector. In specific implementations, word embedding technology or a pre-trained language model can be used to convert each knowledge point description into a 300-dimensional vector representation.

[0073] After conversion to vector form, the system calculates the distance between the vectors of any two knowledge points to obtain the dependency strength value. For example, by calculating cosine similarity, the system determined that the dependency strength between "Linear Regression" and "Decision Tree" is 0.72, while the dependency strength between "Linear Regression" and "Data Preprocessing" is 0.85, indicating that "Linear Regression" and "Data Preprocessing" have a higher degree of correlation. Based on the calculated dependency strength values ​​between all pairs of knowledge points, the system constructs an n×n knowledge dependency matrix M, where n is the total number of knowledge points. Each element M[i][j] in the matrix represents the dependency strength of knowledge point i on knowledge point j.

[0074] The system performs singular value decomposition on the constructed knowledge dependency matrix to obtain eigenvectors of the knowledge points. For example, for the dependency matrix of the "Basic Machine Learning Training" mentioned above, the system obtains five main eigenvectors through singular value decomposition, each corresponding to the position coordinates of a knowledge point in the knowledge space. Using these eigenvectors, the system constructs the distribution structure of the knowledge points in the knowledge space and generates a visual knowledge dependency graph. In this graph, each node represents a knowledge point, and the lines connecting the nodes represent the dependency relationships. The thickness or color of the lines indicates the strength of the dependency.

[0075] In the generated knowledge dependency graph, the system calculates the out-degree and in-degree values ​​of each knowledge point. The out-degree value indicates the degree of influence of the knowledge point on other knowledge points, and the in-degree value indicates the degree of influence of other knowledge points on the knowledge point. For example, the out-degree value of "data preprocessing" is 4.2 and the in-degree value is 1.3, indicating that "data preprocessing" is the foundation of multiple other knowledge points. The system determines the hierarchical position coefficient of the current knowledge point based on the ratio of the out-degree value to the in-degree value. Specifically, the hierarchical position coefficient is equal to the out-degree value divided by the in-degree value. If the value is greater than 1.5, the knowledge point belongs to the basic level; if the value is between 0.7 and 1.5, the knowledge point belongs to the intermediate level; if the value is less than 0.7, the knowledge point belongs to the application level.

[0076] Based on the calculated hierarchical position coefficients, the system stratifies the knowledge points in the knowledge dependency graph. In this embodiment, the hierarchical position coefficient of "data preprocessing" is 3.23, which is classified as the basic level; the hierarchical position coefficient of "linear regression" is 1.2, which is classified as the intermediate level; and the hierarchical position coefficient of "model evaluation" is 0.5, which is classified as the application level. The system groups knowledge points with the same hierarchical position coefficient and a dependency strength value greater than a preset strength threshold into a group, thereby obtaining multiple training subtasks. For example, when the preset strength threshold is set to 0.7, the system divides "data preprocessing" and "feature engineering" into one training subtask, and divides "linear regression" and "decision tree" into another training subtask.

[0077] In practical applications, based on the "Programming Language Training" task requirement, the system extracted knowledge points such as "variable declaration," "data types," "control structures," "function definitions," and "exception handling." After converting these into vector representations, the system calculated the dependency strength between "variable declaration" and "data types" to be 0.92, and the dependency strength between "function definitions" and "exception handling" to be 0.65. After constructing the dependency matrix, singular value decomposition was used to obtain the eigenvectors of each knowledge point, generating a knowledge dependency graph. The calculated out-degree value for "variable declaration" was 5.1, the in-degree value was 0.9, and the hierarchical position coefficient was 5.67, placing it at the foundational level. The out-degree value for "exception handling" was 1.8, the in-degree value was 3.2, and the hierarchical position coefficient was 0.56, placing it at the application level. Ultimately, the system assigned "variable declaration" and "data types" to the first training subtask, "control structures" and "function definitions" to the second, and "exception handling" and "debugging techniques" to the third.

[0078] In this embodiment, by vectorizing the content of knowledge points and calculating the dependency strength value, the association relationship between knowledge points can be accurately identified; the characteristic vectors of knowledge points are extracted through singular value decomposition, and the distribution structure in the knowledge space is constructed to form a knowledge dependency graph with a hierarchical relationship; the hierarchical position coefficient is further calculated according to the in-and-out degree ratio to realize the hierarchical positioning of the knowledge points, and based on this, the knowledge points are hierarchically clustered, and knowledge points with strong correlation and similar hierarchies are divided into the same training subtask, thereby ensuring that the knowledge organization within each subtask is tight and the progression is reasonable, which helps to improve the coherence of the training content and learning efficiency.

[0079] In an optional embodiment, establishing an execution priority sequence for training subtasks, evaluating the difficulty coefficient of each training subtask based on knowledge node coverage, and generating a training task decomposition plan including execution order and difficulty rating include:

[0080] Obtain the knowledge dependency relationship of the training subtask, calculate the number of direct input edges and direct output edges of the knowledge node to obtain the initial dependency value, calculate the number of indirect input edges and indirect output edges of the knowledge node to obtain the transitive dependency value, and calculate the dependency coefficient of the training subtask based on the initial dependency value and the transitive dependency value;

[0081] Marking reachable nodes on the knowledge dependency graph, calculating a node transfer factor according to the hierarchical distribution of the reachable nodes, and taking the product of the node transfer factor and the total number of reachable nodes as the importance value of the training subtask;

[0082] A state transition matrix is ​​constructed using the dependency coefficients and importance values, a state distribution vector is calculated by eigenvalue decomposition, the state distribution vector is used as the execution priority sequence of the training subtasks, the degree centrality and betweenness centrality of the knowledge nodes are calculated, the weighted sum of the degree centrality and betweenness centrality is used as the node coverage, and the node weight is calculated in combination with the reference chain length of the knowledge node;

[0083] The weights of all knowledge nodes in the training subtasks are accumulated to obtain a difficulty coefficient, and the difficulty coefficients of all training subtasks are graded according to the execution priority sequence to generate a training task decomposition plan including an execution order and a difficulty rating.

[0084] In a specific implementation, the knowledge dependencies of the training subtasks are first obtained. For example, a programming training course consisting of 50 knowledge nodes, including fundamental concepts such as "variable declaration," "conditional statements," and "loop structures," is used as an example. The dependencies between the knowledge nodes are analyzed, and the initial dependency value of each node is calculated. For the "variable declaration" node, its number of direct incoming edges is 0 (it depends on no other nodes) and its number of direct outgoing edges is 7 (it is directly dependent on 7 other nodes). Based on this, the initial dependency value of "variable declaration" is calculated to be -7 (incoming edges minus outgoing edges). Simultaneously, the indirect incoming edges of "variable declaration" are calculated to be 0 and the indirect outgoing edges are 22 (it is indirectly dependent on 22 nodes through intermediate nodes), resulting in a transitive dependency value of -22. Combining the initial dependency value and the transitive dependency value, with a weight ratio of 8:2, the dependency coefficient of the "variable declaration" node is calculated to be -9.8, indicating that this node is fundamental to a large amount of other knowledge.

[0085] Marking reachable nodes on the knowledge dependency graph, starting from the "Variable Declaration" node, reveals 29 reachable nodes distributed across four different levels. The first level has 7 directly reachable nodes, the second level has 12 nodes, the third level has 8 nodes, and the fourth level has 2 nodes. The node transfer factor is calculated to be 0.82, which is obtained by calculating the decreasing ratio of the number of nodes at each level. Multiplying the transfer factor of 0.82 by the total number of reachable nodes (29) yields an importance value of 23.78 for the "Variable Declaration" node.

[0086] Using the dependency coefficients and importance values ​​calculated above, a state transition matrix is ​​constructed. For 50 knowledge nodes, the system creates a 50×50 matrix, where each element represents the probability of transitioning from one node to another. For example, the transition probability from "Variable Declaration" to "Conditional Statement" is 0.23, calculated based on the strength of the dependency relationship between the two nodes. The system processes this matrix using the eigenvalue decomposition algorithm to obtain a stable state distribution vector. The elements of this vector represent the relative importance of each knowledge node, with larger values ​​indicating higher priority. Simultaneously, the coverage of the knowledge nodes is calculated. For the "Variable Declaration" node, its degree centrality is 7 (it is directly connected to 7 other nodes) and its betweenness centrality is 285 (it is on 285 shortest paths). Using a weight ratio of 6:4, the node coverage is calculated to be 122.2. Considering the node's citation chain length of 4 (the depth of citations within the document), the calculated weight of the "Variable Declaration" node is 488.8.

[0087] For training subtask A, which contains the three knowledge nodes "Variable Declaration," "Data Type," and "Operator," the cumulative weights of these three nodes, 488.8, 432.5, and 405.2, yield a difficulty coefficient of 1326.5 for subtask A. Similarly, the difficulty coefficients of the other training subtasks are calculated. Based on the execution priority sequence, all training subtasks are categorized into five levels of difficulty: elementary (difficulty coefficient <1000), basic (1000-2000), intermediate (2000-3000), advanced (3000-4000), and expert (>4000). For example, training subtask A is classified as basic.

[0088] The resulting training task decomposition plan includes execution order and difficulty ratings. In this plan, training subtasks are sorted by execution priority and labeled with their corresponding difficulty ratings. For example, the system recommends completing basic-level training subtask A (priority 0.082, difficulty coefficient 1326.5) first, followed by intermediate-level training subtask C (priority 0.075, difficulty coefficient 2245.8), and finally advanced-level training subtask B (priority 0.063, difficulty coefficient 3562.1).

[0089] Through this method, the system can generate a scientific and reasonable training task execution plan based on the dependencies and importance between knowledge nodes, helping learners to master relevant knowledge more effectively. At the same time, the assessment of difficulty coefficients can help teachers or training institutions better allocate teaching resources and provide targeted guidance for learners of different levels. This method is particularly suitable for learning plans for complex knowledge systems, such as programming, engineering design, medical training and other fields. By calculating the initial dependency values ​​and transitive dependency values ​​of knowledge nodes, the direct and indirect relationships between knowledge points are fully reflected, and the dependency coefficients of subtasks are accurately calculated; the importance values ​​are determined by combining the hierarchical distribution of reachable nodes, so that task sorting takes into account both structural hierarchy and dissemination breadth; the subtask priority sequence is constructed through state transition matrix and eigenvalue decomposition to ensure that the training path is scientific and reasonable; at the same time, centrality analysis and reference chain length are introduced to evaluate the weight of knowledge nodes, and finally the task difficulty is quantified, achieving difficulty grading for all subtasks, thereby generating a training task decomposition plan with a clear structure, controllable execution, and adaptive difficulty.

[0090] In an optional embodiment, marking reachable nodes on a knowledge dependency graph, and calculating a node transfer factor according to the hierarchical distribution of the reachable nodes includes:

[0091] Selecting a starting node in a knowledge dependency graph, generating a node access mark sequence, traversing the knowledge dependency graph using a depth-first search, recording the node access depth and access path, marking the target node that can be reached by the starting node as a reachable node, and forming an access record of the reachable node;

[0092] Dividing the node hierarchy by using the access depth of the node in the access record, calculating the dependency weights of adjacent nodes, assigning hierarchical identifiers to the reachable nodes based on the dependency weights and the access depth, and obtaining a hierarchical distribution structure of the nodes;

[0093] Counting the number of preceding and succeeding nodes of the nodes based on the hierarchical distribution structure, calculating the rate of change of the number of nodes in adjacent levels, and calculating the inter-layer connection strength of the nodes based on the rate of change and the number of preceding and succeeding nodes;

[0094] A node diffusion factor is generated based on the inter-layer connection strength, and a node transfer probability table is constructed in combination with the hierarchical distribution structure. The hierarchical influence strength of the node is calculated through the transfer probability table. The node diffusion factor and the hierarchical influence strength are weightedly combined to generate a node knowledge transfer factor.

[0095] In one embodiment, the nodes of the knowledge dependency graph represent knowledge points, and the edges represent the dependency relationships between knowledge points. First, a starting node is selected from the knowledge dependency graph, such as the "data structure" node, and then a node access mark sequence is generated to record whether the node has been visited. The knowledge dependency graph is traversed using a depth-first search algorithm. For each node visited, its access depth and the access path from the starting node to the node are recorded. For example, when accessing the "data structure" node, its access depth is 0, and the access path only contains itself; when continuing to access the "array" node, its access depth is 1, and the access path is "data structure → array". During the traversal process, the target nodes that can be reached by the starting node are marked as reachable nodes, forming an access record of the reachable nodes. For example, starting from the "data structure" node, the reachable nodes include "array", "linked list", "sorting algorithm", etc., and the access depth and access path of each node are recorded.

[0096] Node levels are divided based on the access depth of the nodes in the access records. For example, nodes with an access depth of 0 belong to layer 0, nodes with an access depth of 1 belong to layer 1, and so on. When calculating the dependency weight between adjacent nodes, the edge weights between nodes and the importance of the nodes are considered. If the edge weight from node A to node B is 0.8, it means that node B has a high degree of dependence on node A. Based on the dependency weight and access depth, hierarchical identifiers are assigned to reachable nodes to obtain a hierarchical distribution structure of the nodes. In actual applications, "data structure" nodes may be identified as layer 0, "array" and "linked list" nodes may be identified as layer 1, and "sorting algorithm" nodes may be identified as layer 2.

[0097] Based on the hierarchical distribution structure, count the number of predecessor and successor nodes for each node. A predecessor node refers to the node that points to the current node, and a successor node refers to the node to which the current node points. Taking the "sort algorithm" node as an example, its predecessor nodes may include "array" and "linked list," with a total of 2 predecessor nodes; its successor nodes may include "quick sort" and "merge sort," with a total of 2 successor nodes. Calculate the rate of change in the number of nodes in adjacent levels, for example, from level 1 to level 2. If level 1 has 5 nodes and level 2 has 8 nodes, the rate of change is (8-5) / 5 = 0.6. Based on the rate of change and the number of predecessor and successor nodes, calculate the inter-level connection strength of the node. For example, for a "sort algorithm" node, if its predecessor nodes are 2 and its successor nodes are 2, and the rate of change in the number of nodes in adjacent levels is 0.6, its inter-level connection strength may be 2 × 2 × 0.6 = 2.4.

[0098] Based on the strength of inter-layer connections, a node diffusion factor is generated. The node diffusion factor indicates the ability of knowledge to spread from one node to its neighboring nodes. For example, the diffusion factor of the "Sorting Algorithm" node might be 0.8, indicating strong knowledge dissemination. Based on the hierarchical distribution structure, a node transfer probability table is constructed to record the probability of knowledge transfer from one node to other nodes. For example, the transfer probability from the "Data Structure" node to the "Array" node is 0.7, and the transfer probability to the "Linked List" node is 0.6. Using the transfer probability table, the node's hierarchical influence strength is calculated, indicating the degree of influence a node has on nodes in its own layer and on nodes in other layers. For example, the influence strength of the "Data Structure" node on nodes in the first layer is 0.65, and its influence strength on nodes in the second layer is 0.45. Finally, the node diffusion factor and the hierarchical influence strength are weighted together to generate the node's knowledge transfer factor. If the "Data Structure" node has a diffusion factor of 0.85 and an influence strength on the first layer is 0.65, its knowledge transfer factor might be 0.85 × 0.65 = 0.5525.

[0099] In practical applications, knowledge dependency graphs containing thousands of knowledge points can be processed. For example, a knowledge dependency graph in the field of computer science might include multiple categories such as "Programming Languages," "Algorithms," and "Operating Systems," each of which contains numerous sub-categories. Using the above method, the knowledge transfer factor for each knowledge point is calculated, providing data support for personalized learning path planning. For example, before learning "Algorithms," a learner might be recommended to first study "Data Structures," which has a higher transfer factor, then "Sorting Algorithms," and finally "Quick Sort" and "Merge Sort."

[0100] This node transfer factor calculation method based on the knowledge dependency graph can effectively capture the dependency relationship and transfer characteristics between knowledge points, provide a reliable decision-making basis for adaptive learning systems, knowledge recommendation systems and learning path planning, and improve learning efficiency and knowledge acquisition quality.

[0101] In this embodiment, accurate modeling of the hierarchical positioning and transfer capabilities of knowledge nodes in the graph can be achieved. Access records are generated through depth-first traversal, clarifying the reachability relationship and path depth between nodes and building a clear hierarchical distribution structure. Hierarchical identifiers are assigned using access depth and dependency weights, refining the relative positions between nodes. Furthermore, by statistically analyzing the rate of change in the number of nodes between levels and the number of preceding and following nodes, the strength of inter-layer connections is quantified, reflecting the potential for information dissemination between layers. Combining the diffusion factor with the hierarchical influence strength, a complete knowledge transfer factor evaluation system is formed, providing efficient support for optimizing the dissemination path of training content, evaluating node importance, and formulating task execution strategies.

[0102] In an optional embodiment, based on the running status of the virtual workplace training agent, the knowledge complementarity score, the cooperation proficiency score, and the task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score, including:

[0103] Obtaining the operational status data of the virtual workplace training agent, including the agent's knowledge status data, interaction status data, and resource status data;

[0104] Extracting the knowledge vectors and skill vectors of the agents based on the knowledge state data, calculating the difference distribution of the knowledge structures between the agents, analyzing the coverage of the knowledge transfer paths between the agents, and calculating the knowledge complementarity scores between the agents based on the difference distribution and coverage;

[0105] Extracting collaborative records of agents based on the interaction state data, including interaction delay, interaction duration, and task achievement, introducing a time decay function into the collaborative records, and calculating a cooperation proficiency score between agents based on the collaborative records processed by the time decay function;

[0106] Analyze the computing resource occupancy, storage resource occupancy, and communication resource occupancy of the agent based on the resource status data, evaluate the resource scheduling efficiency and task switching loss of the agent, and calculate the task carrying capacity score of the agent based on the resource scheduling efficiency and task switching loss;

[0107] The task scenario features are extracted, and the weight coefficients of each score are calculated based on the task scenario features. The knowledge complementarity score, cooperation proficiency score, and task carrying capacity score are nonlinearly weighted and fused with the weight coefficients to generate an intelligent agent collaborative score.

[0108] This embodiment provides a method for calculating an agent collaboration score based on the operational status of a virtual workplace training agent. This method first obtains the operational status data of the virtual workplace training agent, including knowledge status data, interaction status data, and resource status data. It then calculates the knowledge complementarity score, coordination proficiency score, and task carrying capacity score, respectively. Finally, it performs a weighted fusion based on the characteristics of the task scenario to generate an agent collaboration score.

[0109] The scoring process is based on obtaining operational status data for virtual workplace training agents. Real-time data on the agent's performance in the virtual workplace environment is collected through an API. Knowledge status data includes the agent's mastered concepts, skill mastery, and knowledge update timestamps. Interaction status data includes inter-agent messaging, collaborative task completion, and feedback. Resource status data includes metrics such as CPU utilization, memory usage, and network bandwidth consumption. For example, for a sales training agent, its knowledge status data might include "Product Knowledge: 0.85, Negotiation Skills: 0.73, Customer Management: 0.91," etc.

[0110] To calculate the knowledge complementarity score, the agent's knowledge vector and skill vector are first extracted from the knowledge state data. A knowledge vector is represented as a point in an n-dimensional space, with each dimension corresponding to the mastery of a knowledge domain. For example, agent A's knowledge vector might be [0.9, 0.5, 0.3, 0.8], representing its knowledge level in four domains. The skill vector reflects the agent's practical ability, such as [0.7, 0.6, 0.9], indicating proficiency in three core skills. The difference distribution between the agents' knowledge structures is calculated using a vector difference metric, such as the complement of cosine similarity or the normalized value of Euclidean distance. For example, if the Euclidean distance between the knowledge vectors of two agents is 1.2, normalization yields a difference of 0.6. The coverage of the knowledge transfer paths between the agents is also analyzed, a knowledge graph is constructed, and the efficiency of knowledge flow is assessed using a path analysis algorithm. For example, if two agents can transfer all knowledge through three intermediate nodes, the coverage rate is 95%. Finally, the knowledge complementarity score is calculated based on the difference distribution and coverage, ranging from 0 to 100.

[0111] To calculate the coordination proficiency score, the agent collaboration records are extracted from the interaction state data. Interaction latency refers to the time interval between a request from one agent and a response from another, such as an average response time of 2.3 seconds. Interaction duration represents the total time required to complete a collaborative task, such as an average collaboration time of 15 minutes. Task achievement measures the quality of completion of task objectives, such as an average achievement rate of 85%. The system introduces a time decay function to collaboration records, giving more recent collaboration records a higher weight. For example, a collaboration record from a month ago might be weighted only 0.7 times as heavily as one from the past week. The coordination proficiency score between agents is calculated based on these weighted collaboration records, taking into account response speed, collaboration efficiency, and task quality. In specific implementations, a base score is calculated using the formula "inverse of latency × 0.3 + inverse of duration × 0.3 + achievement × 0.4." This is then multiplied by the time weight to obtain the final coordination proficiency score, which also ranges from 0 to 100.

[0112] When calculating the task carrying capacity score, the agent's resource usage is analyzed based on resource status data. Computing resource usage includes peak and average CPU usage, such as a peak of 85% and an average of 45%; storage resource usage includes memory usage and storage space requirements, such as a memory usage of 600MB; and communication resource usage includes network bandwidth usage and data transmission volume, such as an average bandwidth usage of 20Mbps. Next, the agent's resource scheduling efficiency is evaluated, including the rationality of resource allocation and parallel processing capabilities, such as resource utilization reaching 75%. Task switching loss reflects the time and resource overhead required for the agent to switch from one task to another, such as an average switching time of 1.2 seconds and a resource loss rate of 8%. The agent's task carrying capacity score is calculated based on resource scheduling efficiency and task switching loss, with a value range of 0-100.

[0113] To generate the final collaboration score, task scenario characteristics are first extracted, including task complexity, time urgency, and resource constraints. For example, a high-complexity, time-critical, and resource-limited scenario might be represented as [0.9, 0.8, 0.7]. Weight coefficients for each score are calculated based on the task scenario characteristics. For example, in a knowledge-intensive task, the knowledge complementarity score might have a weight of 0.5, the coordination proficiency score a weight of 0.3, and the task carrying capacity score a weight of 0.2. The system uses a nonlinear weighted fusion method to combine these three scores with the weight coefficients to generate the final agent collaboration score. Nonlinear fusion considers the interactions between the scores, rather than a simple linear combination. For example, if the knowledge complementarity score is 80, the coordination proficiency score is 70, and the task carrying capacity score is 90, with weights of 0.5, 0.3, and 0.2, respectively, the final collaboration score might be 78. This score can be used to guide agent combination optimization and task allocation decisions in the virtual workplace.

[0114] In this embodiment, by introducing the operating status data of the virtual workplace training agent, including the knowledge status, interaction status and resource status, a more comprehensive and dynamic collaborative capability evaluation mechanism is proposed on the basis of the existing technology that only evaluates the agent's capabilities based on the task completion rate or static characteristics. The existing technology often ignores key factors such as the differences in knowledge structure between agents, the timeliness of the interaction process and the efficiency of resource scheduling, resulting in delayed and lack of pertinence in the collaborative evaluation results. Starting from the complementarity of knowledge, the present application extracts the knowledge and skill vectors of the agent, analyzes its structural differences and transmission path coverage, and improves the collaborative evaluation's perception of the adaptability of the knowledge structure; introduces a time decay function to dynamically adjust the interaction records, and strengthens the characterization of the evolution of cooperation proficiency over time; evaluates task switching loss and scheduling efficiency through resource status data, effectively filling the blind spot of task carrying capacity considerations in the existing evaluation system. Finally, the weights of various indicators are dynamically adjusted in combination with the characteristics of the task scenario to achieve nonlinear fusion of multi-dimensional scores and generate more adaptive and real-time collaborative scoring results. This improvement is based on improving the rationality of collaborative task allocation and the collaborative efficiency of training agents, effectively enhancing the accuracy, flexibility and practical value of the evaluation system.

[0115] In an optional embodiment, a knowledge topology network is established based on the difficulty coefficient in the training task decomposition scheme and the agent collaborative score, the transfer path scores of the agents in the knowledge topology network are extracted, and the agent with the highest transfer path score and whose computing resources are not saturated is selected as the leading training agent, including:

[0116] Extracting the difficulty coefficient of the training subtask from the training task decomposition plan, converting the difficulty coefficient into a node-based weight, and converting the agent collaboration score into a node-dynamic weight;

[0117] An initial knowledge topology network is constructed using the node basic weights and node dynamic weights, the knowledge hierarchy relationship of adjacent nodes in the initial knowledge topology network is analyzed, the knowledge overlap between nodes is calculated, the direction of knowledge flow is determined based on the knowledge hierarchy relationship and the knowledge overlap, and a directed transfer path is constructed;

[0118] Calculating the node complexity on the directed transfer path based on the directed transfer path, substituting the node complexity and the transfer distance into an exponential decay function to obtain a path knowledge decay value, and multiplying the knowledge overlap between nodes by the path knowledge decay value to obtain a transfer path score;

[0119] The task queue length, memory occupancy rate and response delay time of the intelligent agent are collected, and the processing load index is calculated. The processing load index is compared with the preset multi-dimensional saturation threshold to obtain the resource status judgment result of the intelligent agent. The transfer path score and the resource status judgment result are combined and evaluated, and the intelligent agent with the highest transfer path score and which has not reached the resource saturation state is selected as the leading training intelligent agent.

[0120] This embodiment provides a method for selecting the optimal lead training agent by building a knowledge topology network based on the difficulty coefficients in the training task decomposition scheme and the collaborative scoring of agents. The method then extracts the transfer path scores of the agents within the knowledge topology network. By building the knowledge topology network and calculating the transfer path scores, the method combines the resource status of the agents to determine the most suitable candidate for the lead training agent.

[0121] During implementation, the difficulty coefficients of the training subtasks are first extracted from the training task decomposition scheme. For example, for an image recognition training task consisting of multiple subtasks, the difficulty coefficient of subtask A is 0.8, the difficulty coefficient of subtask B is 0.6, and the difficulty coefficient of subtask C is 0.4. These difficulty coefficients are converted into node-based weights through linear normalization, so that the sum of the basic weights of all nodes is 1. Simultaneously, the system obtains the collaborative scores between agents. For example, the collaborative score between agents X and Y is 0.75, and the collaborative score between agents X and Z is 0.6. These collaborative scores are converted into node dynamic weights through the softmax function. Node dynamic weights reflect the quality of the agents' performance in historical collaboration.

[0122] An initial knowledge topology network is constructed using node base weights and node dynamic weights. In this network, each node represents an agent, and the size of the node is determined by the weighted sum of the base weight and the dynamic weight. For example, a combination of a base weight of 0.7 and a dynamic weight of 0.3 is used. Connections between nodes represent the potential for knowledge transfer between agents, and the initial connection strength is determined by the collaboration score between the two agents.

[0123] Analyze the knowledge hierarchy relationships between adjacent nodes in the initial topological network. This step is achieved by comparing the knowledge base complexity of the nodes. Assuming that the knowledge complexity of agent X is 8.5 and that of agent Y is 7.2, then X has a higher knowledge hierarchy than Y. The system calculates the knowledge overlap between the nodes. For example, using the cosine similarity method of word vectors, the knowledge overlap between agents X and Y is 0.65. Based on the knowledge hierarchy relationship and overlap, the direction of knowledge flow is determined, flowing from high-level nodes to low-level nodes, and a directed transfer path is constructed. In this example, the weight of the directed path from X to Y is 0.65 multiplied by the difference in their hierarchical levels (8.5-7.2), that is, 0.845.

[0124] Based on a directed transfer path, calculate the complexity of the nodes on the path. For example, in the path X→Y→Z, the complexity of X is 8.5, Y is 7.2, and Z is 6.8. Substitute the node complexity and transfer distance into the exponential decay function to calculate the path knowledge decay value. Assume that the decay function used is e (-0.1d) , where d is the transfer distance, the attenuation value from X to Y is 0.905, and the attenuation value from Y to Z is 0.905. The total attenuation value for the path X→Y→Z is 0.905 × 0.905 = 0.819. Multiplying the knowledge overlap between nodes by the path knowledge decay value yields the transfer path score. Assuming the overlap from X to Y is 0.65 and the overlap from Y to Z is 0.58, the transfer path score for the path X→Y→Z is 0.65 × 0.58 × 0.819 = 0.309.

[0125] The agent's task queue length, memory utilization, and response latency are collected in real time. For example, agent X's task queue length is 15, memory utilization is 68%, and response latency is 120 milliseconds. These metrics are weighted averaged to calculate the processing load index, resulting in a processing load index of 0.72 for X. The system has preset multi-dimensional saturation thresholds, such as a task queue length threshold of 20, a memory utilization threshold of 85%, and a response latency threshold of 200 milliseconds. Comparing X's processing load index of 0.72 with the standardized value of 0.85 corresponding to the preset thresholds indicates that X has not reached resource saturation.

[0126] A similar calculation is performed for all agents to determine their transfer path scores and resource status. For example, if the path X→Y→Z scores 0.309, X has not reached resource saturation; the path Y→X→W scores 0.278, Y has not reached resource saturation; and the path Z→Y→X scores 0.245, Z has reached resource saturation. By combining the transfer path scores and resource status, the agent with the highest transfer path score and resource saturation status is selected as the lead training agent. In this example, agent X is selected as the lead training agent.

[0127] Through the above steps, the system successfully selected the most suitable leading training agent from multiple agents. This agent has the highest transfer path score in the knowledge topology network, and its computing resources have not yet reached saturation. It can effectively assume the leading role of the training task and improve the efficiency and quality of the entire training process.

[0128] In this embodiment, by integrating training task difficulty with agent collaboration capabilities, a knowledge topology network encompassing static cognitive capabilities and dynamic operational states is established, significantly improving the scientific and adaptable nature of training task allocation. Compared to existing approaches that rely solely on static allocation based on preset rules or task history records, which fail to fully reflect real-time differences between agents and the degree of knowledge structure matching, this application proposes mapping difficulty coefficients and collaboration scores into basic and dynamic weights, respectively, and constructing knowledge flow paths. Transfer directions are determined by analyzing knowledge hierarchies and overlaps, and the effectiveness of each path is accurately quantified by combining path complexity with a knowledge decay model. Real-time metrics such as task queues, resource utilization, and response latency are also introduced to conduct a multi-dimensional assessment of agent operational load, enabling a coordinated assessment of resource status and knowledge transfer efficiency. Ultimately, based on the comprehensive evaluation results, the optimal agent is dynamically selected as the task leader, optimizing the task allocation logic from the source, improving knowledge transfer efficiency and the overall operational performance of the training system. Aiming to achieve efficient training scheduling, this technology balances task adaptability and resource carrying capacity, addressing the issues of single-point allocation results and slow response in existing systems.

[0129] In an optional embodiment, determining the collaborative training agent based on the knowledge distribution of the transfer path and generating the agent combination configuration scheme for the training subtasks includes:

[0130] Extract the transfer path of the leading training agent in the knowledge topology network, calculate the knowledge weight of each node on the path, build a knowledge transfer chain based on the knowledge association strength between nodes, analyze the knowledge coverage based on the knowledge transfer chain, and generate knowledge distribution characteristics;

[0131] Calculate the matching degree between the knowledge distribution characteristics and the training subtask requirements, identify uncovered knowledge content, search for agents with the uncovered knowledge content in the knowledge topology network, calculate the degree of knowledge complementarity between the candidate agent and the leading training agent based on the knowledge association strength, and select the agent with the highest degree of knowledge complementarity as the collaborative training agent;

[0132] Based on the strength of the knowledge association, the order of knowledge transfer between the leading training agent and the collaborative training agent is determined, computing resources are allocated to the agents according to the knowledge distribution characteristics and the degree of knowledge complementarity, and the knowledge transfer order and computing resource allocation results are combined to generate an agent combination configuration plan for the training subtask.

[0133] For example, a knowledge topology network must first be established. This network consists of multiple agent nodes and their connections, with each agent node possessing knowledge distribution in different domains. The weights of the connections between nodes represent the strength of the knowledge association, ranging from 0 to 1, with higher values ​​indicating stronger associations. For example, in an autonomous driving training task, the knowledge association strength between the perception and recognition agent and the decision-making and control agent might be 0.75, while the association strength with the environment simulation agent might be 0.45.

[0134] When extracting the transfer path of the leading training agent in the knowledge topology network, a depth-first search algorithm is used to traverse all nodes directly connected to the agent, recording the path sequence formed during the traversal. Assuming that the leading training agent is A, and its directly connected agents are B, C, and D, with knowledge association strengths of 0.8, 0.6, and 0.4, respectively, the AB path with a strength of 0.8 is searched first, followed by the AC and AD paths.

[0135] Calculating the knowledge weight of each node on a path involves two factors: the node's own knowledge and its position in the path. A node's own knowledge can be represented by a vector. For example, agent A's knowledge is [0.9, 0.7, 0.5, 0.2], corresponding to the four knowledge dimensions. The influence of a node's position in the path decreases as it moves away from the dominant node, with a coefficient of 0.8. For example, on the ABE path, the effective knowledge weights of the three nodes are A's weight, B's weight multiplied by 0.8, and E's weight multiplied by 0.64.

[0136] When constructing a knowledge transfer chain, paths are filtered based on the strength of the knowledge associations between nodes. A threshold of 0.5 is set to retain only paths where the knowledge association strength of all links on the path exceeds the threshold. For example, if the association strength of AB is 0.8 and the association strength of BE is 0.4, the ABE path is pruned, retaining only the AB portion. The resulting knowledge transfer chain may contain multiple paths, such as {AB, AC, BF}.

[0137] Based on the knowledge transfer chain, we analyze knowledge coverage and calculate the cumulative knowledge vectors of all nodes in the chain. Assuming the knowledge vector dimension of an agent is 5, corresponding to five knowledge domains, the knowledge vectors of agents A, B, and C are [0.9, 0.7, 0.5, 0.2, 0.1], [0.2, 0.8, 0.7, 0.6, 0.3], and [0.4, 0.3, 0.9, 0.7, 0.8], respectively. After accounting for position decay, the cumulative knowledge coverage vector of the knowledge transfer chain might be [1.12, 1.27, 1.42, 0.92, 0.73]. This constitutes the knowledge distribution feature.

[0138] When calculating the degree of fit between knowledge distribution characteristics and training subtask requirements, it's necessary to first identify the knowledge requirement vector for the training subtask. For example, if the knowledge requirement vector for a training subtask is [0.8, 0.9, 0.7, 0.8, 0.9], comparing it element-wise with the knowledge distribution characteristic vector yields a knowledge gap vector of [-0.32, -0.37, -0.72, -0.12, 0.17]. Negative values ​​indicate a knowledge surplus in that dimension, while positive values ​​indicate a knowledge gap. In this example, the fifth knowledge dimension has a gap of 0.17, identifying it as uncovered knowledge content.

[0139] When searching for agents with uncovered knowledge in a knowledge topology network, agents with high values ​​in the fifth knowledge dimension are selected as candidates. For example, if there are agents G and H in the network, with fifth-dimensional knowledge values ​​of 0.9 and 0.7, respectively, they become candidates.

[0140] When calculating the degree of knowledge complementarity between the candidate agent and the leading training agent, two aspects are considered: the degree of knowledge complementation for uncovered knowledge and the similarity of the overall knowledge structure. The degree of knowledge complementation for uncovered knowledge is directly compared to the gap value and the knowledge volume of the candidate agent in the corresponding dimension. For example, if G can complement 0.9, exceeding the gap by 0.17, the complement rate is 100%; if H can complement 0.7, the complement rate is 100%. The similarity of the overall knowledge structure is calculated using cosine similarity. Assume that the similarity between G and leading training agent A is 0.65, and the similarity between H and A is 0.82.

[0141] Taking all factors into consideration, the degree of knowledge complementarity is calculated as the weighted sum of the complementation rate and the similarity. The weights can be set to 0.7 for the complementation rate and 0.3 for the similarity. G's degree of knowledge complementarity is 0.7 × 1 + 0.3 × 0.65 = 0.895, while H's is 0.7 × 1 + 0.3 × 0.82 = 0.946. H has a higher degree of knowledge complementarity, so H is selected as the collaborative training agent.

[0142] To determine the order of knowledge transfer, a directed graph is constructed based on the strength of knowledge associations. If the knowledge association strength between the leading training agent A and the collaborating training agent H is 0.6 and A possesses the prerequisite knowledge, the transfer order is A → H. Conversely, if H possesses the prerequisite knowledge, the transfer order is H → A. In complex situations, a transfer chain such as A → C → H may form.

[0143] When allocating computing resources based on knowledge distribution characteristics and the degree of knowledge complementarity, the resource allocation ratio is proportional to the importance of the agent in knowledge transfer. Assuming a total of 100 units of computing resources, the leading training agent A, due to its control over the overall training direction, receives 60 units of basic resources. Collaborative training agent H, based on its knowledge complementarity of 0.946, receives 60 × 0.946 = 56.76 units of resources, rounded to 57 units.

[0144] The final training subtask agent combination configuration scheme includes: leading training agent A and collaborative training agent H, the knowledge transfer order is A→H, and the computing resource allocation is 60 units for A and 57 units for H, which are used to complete the knowledge and skills transfer of specific training subtasks.

[0145] In this embodiment, by introducing a comprehensive analysis of knowledge distribution characteristics and knowledge complementarity between agents, the configuration rationality and collaborative efficiency of training subtasks in a multi-agent system are improved. Existing technologies mostly use fixed or empirically set agent combinations, which cannot dynamically adapt to changes in knowledge structure and individual ability differences, resulting in insufficient coverage of some training tasks or uneven resource allocation. This solution extracts the knowledge transfer path of the dominant training agent and constructs a knowledge distribution feature model that includes knowledge weights and correlation strengths to achieve high-precision matching between task knowledge requirements and agent knowledge reserves, thereby identifying knowledge blind spots and intelligently selecting collaborative agents. At the same time, the collaborative process is optimized based on knowledge complementarity and transfer order, and computing resources are reasonably allocated in combination with the knowledge distribution structure to achieve a dual balance between task carrying and knowledge collaboration. This technology effectively solves the problems of random collaborative relationships, resource waste, and weak coverage capabilities in traditional methods, and enhances the adaptability of training tasks and the collaborative capabilities of agents.

[0146] A second aspect of an embodiment of the present invention provides a virtual workplace training task allocation system based on multi-agent collaboration, the system comprising:

[0147] The first unit is used to collect training task requirements and the operating status of the virtual workplace training agent. Based on the training task requirements, it constructs a knowledge dependency graph, identifies the hierarchical relationships and correlations between training knowledge nodes, and divides the training task into multiple training subtasks. It also establishes an execution priority sequence for the training subtasks, assesses the difficulty coefficient of each training subtask based on the knowledge node coverage, and generates a training task decomposition plan that includes the execution order and difficulty rating.

[0148] The second unit is used to calculate the knowledge complementarity score, cooperation proficiency score, and task carrying capacity score between agents based on the operating status of the virtual workplace training agents, and to weight and fuse these scores to generate an agent collaboration score;

[0149] The third unit is used to establish a knowledge topology network based on the difficulty coefficient and agent collaboration score in the training task decomposition plan, extract the transfer path scores of the agents in the knowledge topology network, select the agent with the highest transfer path score and whose computing resources are not saturated as the leading training agent, and determine the collaborative training agents based on the knowledge distribution of the transfer path to generate the agent combination configuration plan for the training subtasks;

[0150] The fourth unit is used to execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

[0151] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0152] processor;

[0153] a memory for storing processor-executable instructions;

[0154] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0155] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0156] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A virtual workplace training task allocation method based on multi-agent collaboration, characterized by: include: Collect training task requirements and the operating status of the virtual workplace training agent, build a knowledge dependency graph based on the training task requirements, identify the hierarchical relationship and correlation between training knowledge nodes, and divide the training task into multiple training subtasks; Establish an execution priority sequence for training subtasks, evaluate the difficulty coefficient of each training subtask based on knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating; Based on the running status of the virtual workplace training agents, the knowledge complementarity score, cooperation proficiency score and task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score; Based on the difficulty coefficient and agent collaboration scores in the training task decomposition scheme, a knowledge topology network is established. The transfer path scores of the agents in the knowledge topology network are extracted. The agent with the highest transfer path score and whose computing resources are not saturated is selected as the leading training agent. The collaborative training agents are determined based on the knowledge distribution of the transfer path, and the agent combination configuration scheme for the training subtasks is generated. Execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

2. The method according to claim 1, characterized in that Based on the training task requirements, a knowledge dependency graph is constructed to identify the hierarchical relationship and degree of association between training knowledge nodes. The training task is then divided into multiple training subtasks, including: Extracting knowledge points from training task requirements, converting the knowledge points into vector form, calculating the distance between the vectors of any two knowledge points to obtain a dependency strength value, and constructing a knowledge dependency matrix based on the dependency strength value; Performing singular value decomposition on the knowledge dependency matrix to obtain eigenvectors of the knowledge points, using the eigenvectors to construct a distribution structure of the knowledge points in the knowledge space, and generating a knowledge dependency graph; In the knowledge dependency graph, the out-degree value and the in-degree value of each knowledge point are calculated, and the hierarchical position coefficient of the current knowledge point is determined according to the ratio of the out-degree value to the in-degree value; The knowledge points in the knowledge dependency graph are layered according to the hierarchical position coefficients, and the knowledge points with the same hierarchical position coefficients and dependency strength values ​​greater than a preset strength threshold are divided into a group to obtain a plurality of training subtasks.

3. The method according to claim 1, characterized in that Establish an execution priority sequence for training subtasks, evaluate the difficulty coefficient of each training subtask based on knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating. Obtain the knowledge dependency relationship of the training subtask, calculate the number of direct input edges and direct output edges of the knowledge node to obtain the initial dependency value, calculate the number of indirect input edges and indirect output edges of the knowledge node to obtain the transitive dependency value, and calculate the dependency coefficient of the training subtask based on the initial dependency value and the transitive dependency value; Marking reachable nodes on the knowledge dependency graph, calculating a node transfer factor according to the hierarchical distribution of the reachable nodes, and taking the product of the node transfer factor and the total number of reachable nodes as the importance value of the training subtask; A state transition matrix is ​​constructed using the dependency coefficients and importance values, a state distribution vector is calculated by eigenvalue decomposition, the state distribution vector is used as the execution priority sequence of the training subtasks, the degree centrality and betweenness centrality of the knowledge nodes are calculated, the weighted sum of the degree centrality and betweenness centrality is used as the node coverage, and the node weight is calculated in combination with the reference chain length of the knowledge node; The weights of all knowledge nodes in the training subtasks are accumulated to obtain a difficulty coefficient, and the difficulty coefficients of all training subtasks are graded according to the execution priority sequence to generate a training task decomposition plan including an execution order and a difficulty rating.

4. The method according to claim 3, characterized in that Marking reachable nodes on the knowledge dependency graph and calculating the node transfer factor according to the hierarchical distribution of the reachable nodes include: Selecting a starting node in a knowledge dependency graph, generating a node access mark sequence, traversing the knowledge dependency graph using a depth-first search, recording the node access depth and access path, marking the target node that can be reached by the starting node as a reachable node, and forming an access record of the reachable node; Dividing the node hierarchy by using the access depth of the node in the access record, calculating the dependency weights of adjacent nodes, assigning hierarchical identifiers to the reachable nodes based on the dependency weights and the access depth, and obtaining a hierarchical distribution structure of the nodes; Counting the number of preceding and succeeding nodes of the nodes based on the hierarchical distribution structure, calculating the rate of change of the number of nodes in adjacent levels, and calculating the inter-layer connection strength of the nodes based on the rate of change and the number of preceding and succeeding nodes; A node diffusion factor is generated based on the inter-layer connection strength, and a node transfer probability table is constructed in combination with the hierarchical distribution structure. The hierarchical influence strength of the node is calculated through the transfer probability table. The node diffusion factor and the hierarchical influence strength are weightedly combined to generate a node knowledge transfer factor.

5. The method according to claim 1, wherein Based on the running status of the virtual workplace training agent, the knowledge complementarity score, cooperation proficiency score and task carrying capacity score between agents are calculated, and the scores are weighted and integrated to generate the agent collaboration score, including: Obtaining the operational status data of the virtual workplace training agent, including the agent's knowledge status data, interaction status data, and resource status data; Extracting the knowledge vectors and skill vectors of the agents based on the knowledge state data, calculating the difference distribution of the knowledge structures between the agents, analyzing the coverage of the knowledge transfer paths between the agents, and calculating the knowledge complementarity scores between the agents based on the difference distribution and coverage; Extracting collaborative records of agents based on the interaction state data, including interaction delay, interaction duration, and task achievement, introducing a time decay function into the collaborative records, and calculating a cooperation proficiency score between agents based on the collaborative records processed by the time decay function; Analyze the computing resource occupancy, storage resource occupancy, and communication resource occupancy of the agent based on the resource status data, evaluate the resource scheduling efficiency and task switching loss of the agent, and calculate the task carrying capacity score of the agent based on the resource scheduling efficiency and task switching loss; The task scenario features are extracted, and the weight coefficients of each score are calculated based on the task scenario features. The knowledge complementarity score, cooperation proficiency score, and task carrying capacity score are nonlinearly weighted and fused with the weight coefficients to generate an intelligent agent collaborative score.

6. The method according to claim 1, characterized in that Based on the difficulty coefficient in the training task decomposition plan and the agent collaborative score, a knowledge topology network is established. The transfer path scores of the agents in the knowledge topology network are extracted. The agents with the highest transfer path scores and whose computing resources are not saturated are selected as the leading training agents. Extracting the difficulty coefficient of the training subtask from the training task decomposition plan, converting the difficulty coefficient into a node-based weight, and converting the agent collaboration score into a node-dynamic weight; An initial knowledge topology network is constructed using the node basic weights and node dynamic weights, the knowledge hierarchy relationship of adjacent nodes in the initial knowledge topology network is analyzed, the knowledge overlap between nodes is calculated, the direction of knowledge flow is determined based on the knowledge hierarchy relationship and the knowledge overlap, and a directed transfer path is constructed; Calculating the node complexity on the directed transfer path based on the directed transfer path, substituting the node complexity and the transfer distance into an exponential decay function to obtain a path knowledge decay value, and multiplying the knowledge overlap between nodes by the path knowledge decay value to obtain a transfer path score; The task queue length, memory occupancy rate and response delay time of the intelligent agent are collected, and the processing load index is calculated. The processing load index is compared with the preset multi-dimensional saturation threshold to obtain the resource status judgment result of the intelligent agent. The transfer path score and the resource status judgment result are combined and evaluated, and the intelligent agent with the highest transfer path score and which has not reached the resource saturation state is selected as the leading training intelligent agent.

7. The method according to claim 1, characterized in that The collaborative training agent is determined based on the knowledge distribution of the transfer path, and the agent combination configuration scheme for generating training subtasks includes: Extract the transfer path of the leading training agent in the knowledge topology network, calculate the knowledge weight of each node on the path, build a knowledge transfer chain based on the knowledge association strength between nodes, analyze the knowledge coverage based on the knowledge transfer chain, and generate knowledge distribution characteristics; Calculate the matching degree between the knowledge distribution characteristics and the training subtask requirements, identify uncovered knowledge content, search for agents with the uncovered knowledge content in the knowledge topology network, calculate the degree of knowledge complementarity between the candidate agent and the leading training agent based on the knowledge association strength, and select the agent with the highest degree of knowledge complementarity as the collaborative training agent; Based on the strength of the knowledge association, the order of knowledge transfer between the leading training agent and the collaborative training agent is determined, computing resources are allocated to the agents according to the knowledge distribution characteristics and the degree of knowledge complementarity, and the knowledge transfer order and computing resource allocation results are combined to generate an agent combination configuration plan for the training subtask.

8. A virtual workplace training task allocation system based on multi-agent collaboration, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect training task requirements and the operating status of the virtual workplace training agent, build a knowledge dependency graph based on the training task requirements, identify the hierarchical relationship and correlation between training knowledge nodes, and divide the training task into multiple training subtasks; Establish an execution priority sequence for training subtasks, evaluate the difficulty coefficient of each training subtask based on knowledge node coverage, and generate a training task decomposition plan that includes execution order and difficulty rating; The second unit is used to calculate the knowledge complementarity score, cooperation proficiency score, and task carrying capacity score between agents based on the operating status of the virtual workplace training agents, and to weight and fuse these scores to generate an agent collaboration score; The third unit is used to establish a knowledge topology network based on the difficulty coefficient and agent collaboration score in the training task decomposition plan, extract the transfer path scores of the agents in the knowledge topology network, select the agent with the highest transfer path score and whose computing resources are not saturated as the leading training agent, and determine the collaborative training agents based on the knowledge distribution of the transfer path to generate the agent combination configuration plan for the training subtasks; The fourth unit is used to execute training subtasks according to the agent collaborative training plan, record the training process data, update the agent collaborative score based on the training process data, and optimize the dynamic task allocation mechanism.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images